40-98%
Lower costs
Up to 98% fewer tokens sent to the LLM, with no signal loss on the content that actually matters.
about · 01
Why Distill?
Every token counts. Distill compresses LLM context upstream — before it ever enters memory — to cut costs, speed up responses, and sharpen output quality.
problem · 02
Noise kills signal.
When you work with an AI coding assistant, you constantly send large context blocks: build outputs, logs, code files, stacktraces. Most of it is redundant or useless.
A typical build error output is thousands of tokens of noise for 5–10 actually useful lines. You're paying for dead weight — and drowning the LLM in context that prevents it from focusing.
Distill fixes this by compressing your context intelligently before it reaches the model. You keep only the signal.
product · 03
One MCP server. Three tools.
Distill is an open-source MCP (Model Context Protocol) server that exposes three always-loaded tools inside Claude Code.
- auto_optimize01 / 03
Detects content type (build, logs, diffs, code, stacktraces) and applies content-aware compression.
- smart_file_read02 / 03
Reads AST structure instead of raw file content. 7 languages, 5 modes (auto, full, skeleton, extract, search).
- code_execute03 / 03
Runs TypeScript in a QuickJS sandbox to batch 5–10 operations in a single MCP call.
No API keys. No cloud services. No auth. Install and start using it immediately.
results · 04
What you get.
lower latency
Faster responses
Less context = less tokens to process = shorter time-to-first-token.
higher signal
Sharper results
Less noise, more signal. The LLM focuses on what matters, output quality goes up.
MCP stdio
Native Claude Code integration
[DISTILL:COMPRESSED] marker, PreCompact hook, distill-compressor subagent, slash commands. Zero config on the API side.
get started
Ready to optimize?
One command to set up Claude Code with Distill.
$bunx distill-mcp setup